Industries / Mining and materials / Glass and ceramics
Inspection
Inspection is checking finished glass and ceramic pieces for cracks, bubbles, stones, wrong size or wrong color, and removing the bad ones before they ship.
5
Milestones achieved, from 1945 to 2016.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1945
Machines look for cracks
Hartford-Empire, the US firm later named Emhart, introduced its HE-74 check inspector in 1945 and its HE-127 automatic finish check inspector in 1954, by Emhart's own timeline. A check is a small crack, and the finish is a bottle's mouth, where the cap seals. In the early 1970s Powers Machinery and Ball sold electro-optical detectors for fine cracks in the finish, and Ball and Owens-Illinois built television-based inspection of bottle sidewalls, according to the machine-vision writer Nello Zuech. By 2002, he wrote, camera systems at glass plants could check up to 800 bottles a minute, though handling limits kept most lines under 600.
Credit
Hartford-Empire (Emhart); Powers Machinery; Ball; Owens-Illinois
1948
Light through empty bottles
RCA received a development contract in 1948 for an electro-optical inspector for returnable bottles. A photomultiplier, a very sensitive light sensor, measured light passing through the bottom of each empty bottle, and any change from the normal level flagged debris or a crack. By 1953 RCA had installed more than 3,000 of these systems on beverage filling lines, mostly in North America, by Zuech's account.
Credit
RCA
c. 1960
Breaking glass on purpose
Tempered glass, which is heated and quenched to make it stronger, can shatter by itself months or years after installation when a tiny nickel sulfide inclusion slowly changes crystal form and expands inside the stressed core. The heat soak test, invented in about 1960 by the account of Andreas Kasper (2018), holds tempered panes at high temperature so that panes carrying such inclusions break in the oven instead of on a building. Kasper gives a commonly quoted oven breakage rate of one pane per six tons of float glass. The test is defined in European standard EN 14179-1.
Credit
not named in the source; standardized by CEN, the European standards body
1998
Cameras grade tiles
A ten-author team including Josef Kittler, Maria Petrou and Francesco De Natale published ASSIST, an automatic system for surface inspection and sorting of tiles, in the Journal of Materials Processing Technology in 1998. Machines now sold for this work, such as SACMI's Flawmaster+, check tiles and slabs for surface, breakage and decoration defects, shade, color, size and squareness, by the maker's brochure.
Credit
C. Boukouvalas, J. Kittler, M. Petrou, F. De Natale, G. Vernazza and colleagues (ASSIST)
2016
Heat soak made cooler
A research project at the SIMaP institute of Grenoble University, coordinated by Andreas Kasper, found in the PhD work of O. Yousfi that above 280 °C, the low end of the holding temperature then set in EN 14179-1, some nickel sulfide compositions cannot finish the change in crystal form the test relies on, so dangerous panes could pass. The 2016 revision of the standard lowered the holding temperature to 260 ± 10 °C.
LessonA screening test is only as good as the physics behind its settings, and this one was set too hot until a study of the inclusions' chemistry showed it.
Credit
O. Yousfi and Andreas Kasper (SIMaP, Grenoble); CEN
Watch for
AI sidewall inspection for bottles
Tiama, a French maker of inspection machines, says deep-learning models in its MCAL 4 AI sidewall machines tell real defects from harmless marks such as mold seams and small blisters, cut job-change time by more than 50%, and raise productivity by up to 1%. It says a furnace with 10 such machines packs 6,500,000 more bottles a year.
Achieved whena glassmaker or outside lab publishes before-and-after false-reject and missed-defect rates for named production lines, checked by re-inspecting samples of rejected and accepted bottles by hand.
Claimed by
Tiama, October 2025
Open
01
Open bottle defect benchmark
A public set of images from production inspection of glass containers (sidewall, base and finish), each labeled with the defect type and its severity on a published scale, with the physical bottles kept so labels can be rechecked, plus a scoring script that reports any model's false-reject rate (good bottles thrown out) at zero missed critical defects.
The briefCredit · unclaimed02
A shared tile shade test
A public set of photographs of fired tiles taken under controlled light, each with a spectrophotometer color reading and the shade group given by several trained graders working apart, plus a test that scores any automatic shade sorter or model against the graders' majority.
The briefCredit · unclaimed
Open 01 · the brief
Open bottle defect benchmark
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of images from production inspection of glass containers (sidewall, base and finish), each labeled with the defect type and its severity on a published scale, with the physical bottles kept so labels can be rechecked, plus a scoring script that reports any model's false-reject rate (good bottles thrown out) at zero missed critical defects.
Why it is still open
Tiama says it trained its models on images from its own fleet of machines, which are not public, and its gains are its own figures. The public sets found are small: Claypo, Jaiyen and Hanskunatai (2023) used 571 body, 570 neck and 120 rotating-bottle images, and the other public sets found in searches of Roboflow and Hugging Face are small student projects. MVTec AD, a widely used benchmark for defect detection, includes a bottle category among its 15 object and texture categories, but only for non-commercial use and not drawn from glass production lines. None found labels severity. The US Department of Agriculture's visual aid already sorts glass container defects into critical, major and minor, so a severity scale exists to label against.
Built with what exists
A turntable rig with an industrial machine-vision camera and LED backlight, the USDA defect definitions, CVAT or Label Studio for labeling, and the open Anomalib library of defect-detection models in PyTorch.
First step
Ask one container plant for a crate of bottles from its reject conveyor, which go back to the furnace as cullet anyway, and photograph 200 of them on a turntable.
Done looks like
At least 20,000 images of at least three glass colors (clear, amber, green) from two or more plants, including at least 1,000 defective bottles whose labels two trained inspectors agree on, released under an open license with the scoring script and a baseline result.
Open 02 · the brief
A shared tile shade test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of photographs of fired tiles taken under controlled light, each with a spectrophotometer color reading and the shade group given by several trained graders working apart, plus a test that scores any automatic shade sorter or model against the graders' majority.
Why it is still open
Shade sorting is sold as a machine function (SACMI's Flawmaster+ brochure lists shade and color checks using RGB and LAB color patterns), but the public tile sets found label cracks, chips and deformations, not shade. A 2025 set on Mendeley Data, for example, has 1,600 images of normal and defective tiles. Saku Kukkonen, H. Kälviäinen and J. Parkkinen compared machine classification of tile shades with human vision in 2001, on five shade classes of one brown tile in a laboratory, but no shared test set with graders' calls was found.
Built with what exists
A color-calibrated camera or flatbed scanner with a color reference chart, a handheld spectrophotometer, Python with the colour-science and scikit-learn libraries, and Label Studio.
First step
Buy two boxes from each of five production lots of one tile design and photograph every tile under the same fixed light.
Done looks like
At least 3,000 tiles from at least ten production lots of three or more designs, each graded by at least three graders, with the agreement among graders (Fleiss' kappa) and a model's agreement with the majority both published, tested on lots the model never saw.